AI-powered music analytics: predicting and optimizing popularity with explainable machine learning

This study develops an AI-powered framework to predict, interpret, and prescribe music popularity in streaming markets using modifiable song attributes. Using 232,442 Spotify tracks with audio features and genre labels, we estimate a random forest regression model with Bayesian hyperparameter tuning to predict Spotify’s Popularity index. We then apply SHAP and LIME to generate multi-level feature attributions (global, genre-, artist-, and song-specific), revealing when aggregate “global drivers” obscure segment-conditional effects consistent with segmentation theory. Building on these explanations, we employ constrained Bayesian optimization to recommend feasible audio-attribute adjustments that raise predicted popularity while respecting user-defined and genre-appropriate bounds reflecting creative intent and production feasibility. It provides actionable decision support for artists, producers, and platforms seeking guidance for content design and targeting. The framework contributes theoretically by operationalizing preference heterogeneity under an aggregate platform metric and methodologically by linking explainable machine learning with constraint-aware prescriptive optimization for feature-level design and promotion.